I am doing research on physics based character control which involves running lots of simulations with articulations and analysing the results, then tweaking the simulations and repeating. Simulating on the CPU, I can run several simulations in parallel, which speeds things up a lot, but I still have so many iterations to run that it can take hours to complete. I have noticed that recent versions of PhysX include CUDA support, and I am wondering whether I can speed things up further by using the GPU, but I have some questions:
what kind of performance might I expect for simulating articulations (containing about 20 links) on GPU vs CPU, assuming that I can run a large number of independent simulations in parallel?
if the GPU is likely to be significantly faster, what would be the best way to structure my problem? For example, would one scene with lots of articulations, or many separate scenes be a better way to go?
can anyone point me to a tutorial or some example code using the GPU in PhysX? I’ve looked at the pages on the Nvidia developer website, but I don’t really feel confident yet that I could get it working easily. Some example code would be really useful.
would the output of the simulation using the GPU be identical to that from the CPU?
I appreciate any help anyone can give me with this.